Self-Configuring Convolutional Neural Network using Evolutionary Algorithm

Abstract

One of the key drivers of today’s AI revolution is an improvement in the field of deep learning. The evolution of deep learning along with big data and the improvement in the hardware industry impacted the Artificial Intelligence industry like never before. As time is going on this field of deep learning is getting improved and we are finding different problems that can be solved to improve it even further. One of them is the structure of a Neural Network. The structure of the neural net must be predefined. Which is done by humans where human error is more likely to occur. Also, it takes a lot of time to form a perfect structure using trial and error method. In this work, we tried to solve this problem using Evolutionary methods. We applied the Genetic Algorithm to find the best architecture of neural network automatically. We used the Genetic Algorithm in Convolutional Neural Networks to create structures. We applied them in two datasets, first the Fashion Mist dataset and then PataNet dataset. The results were interesting. The model provided by the Genetic algorithm performed better than the predefined human structure. Also, the genetic algorithm gave some unusual structure which performed better than any human models.